Does Letting AI Think More Improve How It Picks Stocks?
We gave the same AI model a larger reasoning budget on 100 earnings reports. The simulated account met its preset target, but the 17 changed decisions and a comparison with equal numbers of picks did not show better stock selection.
If you give an AI that reads earnings reports more time to think, does it also get better at choosing which stocks to buy? In a comparison using the same 100 earnings reports from Q2 2026, the higher-reasoning setting left a simulated account worth $244.81 more after trading costs. This account models trades without investing real money.
However, when the number of purchase candidates was held equal, the comparison did not show better stock selection. This article looks separately at the simulated account's results and the accuracy of stock selection.
These results are not actual investment performance. The study covers 1 quarter and 100 reports, and the purchase decision changed in 17 of them. The same result may not appear in other periods.
What was held fixed and what was changed
The conditions for the simulation and the pass criterion were set before the results were seen.
- From 439 earnings reports in Q2 2026, 100 were chosen using a method decided in advance.
- The AI model, the materials, and the instructions are the same. Each report was read once at the usual medium setting and once at a high setting that increases the amount of thinking.
- The simulated account starts with $10,000, can hold up to 8 stocks at the same time, holds each for 10 trading days, and deducts trading costs. Only the 100 reports are switched to the high setting; the remaining 339 stay at the usual setting.
- The pass condition was that the final value of the high-setting account would exceed that of the usual setting by at least $100, which is 1% of the starting capital.
The simulated account improved, but the average for the stocks whose decision changed went the other way
The final value was $13,001.90 at the usual setting and $13,246.71 at the high setting. The difference is +$244.81, which exceeds the +$100 criterion.
Next, the cases where the setting changed the decision were examined. Below are, over the 10 trading days from the starting point defined in the study, the after-cost stock return and its gap versus SPY, an ETF that tracks the S&P 500. The gap is read in percentage points.
| Group | Count | Mean (pt) | Median (pt) |
|---|---|---|---|
| Changed to buy | 13 | −1.15 | −3.90 |
| Dropped from buy | 4 | +1.76 | −0.73 |
| Buy in both settings | 21 | +6.00 | — |
The 13 reports newly turned into buys had a negative mean, and the 4 dropped from buys had a positive mean. However, the median of those 4 dropped was negative. The mean alone does not show that every case moved the same way.
What happens when purchase candidates are held to 25
The usual setting marked 25 reports as buys. Taking the top 25 by score at the high setting as well and comparing the subsequent results gives the following.
The SPY gap is the average gap 10 trading days later for the top 25. The rank correlation shows, across all 100 reports, the link between the rank of the score and the rank of the excess return 10 trading days later.
| Mean SPY gap (pt) | Rank correlation | |
|---|---|---|
| Usual setting | +5.32 | 0.181 |
| High setting | +3.01 | 0.142 |
Among these 100 reports, the high setting's top 25 had a lower average than the usual setting's, and the rank correlation between score and outcome also weakened. A simulated account buying only these 25 ends at $12,948.55, about $53 less than the usual setting.
On the other hand, when the decision on whether to buy was fixed to the usual setting and only the scores were changed to the high setting, the account difference was about +$167. Changes in the order in which candidates that overlapped on the same day were handled, and in which stocks could be held and when they were bought, were tied to most of the original +$244.81 difference. The account difference alone cannot establish that better stocks were distinguished more effectively.
What changed was the interpretation of the stock price, not the calculations
Reading the responses for the 17 reports where the decision changed side by side, the calculations of sales, profit, and price multiples were almost the same. The table below is a tally across all 100 reports read at the two settings. "Undervalued" and "Unclear" are categories for how the AI viewed the stock price; they do not establish a fair price.
| Item | Usual setting | High setting |
|---|---|---|
| Average score | 49.0 | 48.8 |
| Unclear | 22 reports | 11 reports |
| Undervalued | 30 reports | 38 reports |
Of the 13 reports newly turned into buys at the high setting, the price assessment changed from "Unclear" to "Undervalued" in 8.
The next table shows the median price move for each group. "Market open" shows the reaction at the open immediately after the earnings announcement.
| Group | 20 days before announcement | Market open |
|---|---|---|
| Changed to buy (13) | −6.7% | +3.1% |
| Dropped from buy (4) | +14.4% | +6.1% |
The high setting tended to buy stocks that had fallen before the announcement, interpreting the low multiple as not yet pricing in a recovery, and to drop stocks that had risen before the announcement, interpreting growth as already priced in. In this quarter, this contrarian reading did not translate into price moves over 10 trading days. This is an observation, not a comparison that establishes the cause.
Is the usual setting's "Undervalued" rating also weak?
The 2,331 earnings analyses for Q2 2026 produced at the usual setting were also tallied separately. The measurement is the same, but the set differs from the 100 reports used to compare settings.
| Rating | Total count | Mean (pt) | Median (pt) |
|---|---|---|---|
| Undervalued | 210 | +4.14 | +2.85 |
| Fairly valued | 1,399 | −1.90 | −2.18 |
| Overvalued | 152 | −1.25 | −2.51 |
| Unclear | 570 | +1.66 | +0.52 |
The count column is the total number in each rating category, which differs from the number for which a mean could be calculated. The 10-day SPY gap could be confirmed for 207/210 Undervalued, 1,391/1,399 Fairly valued, 152/152 Overvalued, and 567/570 Unclear, for a total of 2,317/2,331. The 14 missing cases were not added to the averages as a profit/loss of 0.
Of the usual setting's 179 buys, 173 were rated "Undervalued." For the 170 of those where the 10-day SPY gap could be confirmed, the mean is +4.75 percentage points and the median is +3.05 percentage points. Buys with a rating other than "Undervalued" numbered 3 or 0 for each rating, which is too few to compare whether the difference in rating produced a difference in performance.
At the high setting, buys increased from 25 to 34. Even though the score cutoff for a buy is the same, more borderline cases were read as undervalued. This change may be related to performance, but the effects of earnings quality, the reaction to the announcement, and the price assessment have not been separated.
These 2,331 reports are a separate sample, from the same quarter read at the usual setting. They are read separately from the comparison of 100 reports in which the amount of thinking was changed.
How outside research relates to this observation
A paper by Anthropic researchers, “Inverse Scaling in Test-Time Compute” (2025), examines cases where increasing the amount of thinking did not improve answers and made them worse, for example through irrelevant reasoning.
Here, the additional reasoning showed up not in the calculation of earnings figures but in the judgment about what is already priced into the stock. Outside research offers a clue for thinking about this, but an observation of 1 quarter and 100 reports is not treated as proof of the same mechanism.
What to confirm before adopting it for investing
The high setting makes the AI think roughly 6 times as much, and costs also rise. The simulated account's results alone cannot tell us whether that cost is justified.
Next, the plan is to read the Q3 2026 earnings reports at the high setting as well, recording the results without using them for trading. In late December, the plan is to check both whether the simulated account improved by at least 1% of starting capital and whether the 10-day SPY gap for the cases where the decision changed is positive, and then reconsider adoption. The results for the cases where "Unclear" changed to "Undervalued" will also be checked. This is the plan as of the time of the study, not a report that the verification has been completed.
What can be read from this
The pass criterion set in advance was the final value of the simulated account. The 17 changes and the 10-day SPY gap are shown as diagnostics afterward. A result from 1 quarter and a small number of changes cannot determine which setting is generally better or why.
When reading about a change in an AI's settings, the takeaway from this study is to check separately the account result, the selection with the number of purchases held equal, and what changed the decision. This article is not investment advice, and past simulations do not guarantee future profits.